Mixtures of large-scale dynamic functional brain network modes

•We propose a new generative model approach for unsupervised inference of dynamic brain networks, known as Dynamic Network Modes (DyNeMo).•DyNeMo avoids making the assumption of mutual exclusivity used in state-based models, such as the Hidden Markov Model (HMM).•DyNeMo reveals a plausible and physiologically meaningful description of resting-state and task MEG data.•DyNeMo is capable of learning longer range dependencies than the HMM, via a recurrent neural network.•We show DyNeMo provides a modest improvement in modelling dynamic spectral features compared to the HMM.